Validation

AV-DSR

A variant of DSR. The same standard DSR, then the 5th percentile after a block resample.

Why

One DSR is a number on one sample. The same series of per-fill net percents is resampled in blocks. Each resample computes a DSR again. The poorer band is the number this variant looks at.

Formula

Which series is read

The per-fill net percents on the discovery-sample user book, the same series as DSR. No second book is opened. The sealed sample is not included.

r = this fill's cash / equity before this fill
SR = mean(r) / std(r)
DSR = Φ( (SR − SR₀) × √(T − 1) / √(1 − γ₃×SR + (γ₄ − 1)/4 × SR²) )

SR₀, the trial count N, the skewness γ₃, and the kurtosis γ₄ are the same as on the DSR page. The denominator of std is T − 1. T is the number of fills.

Resample

Cut this series of r into non-overlapping blocks in time order
Draw those blocks with replacement and assemble a new sample
Recompute one DSR on that sample
Do this B times, giving DSR₁ … DSR_B

AV-DSR = the 5th percentile of these B numbers

What is resampled is the DSR, not the cash. The Monte Carlo that is kept resamples the cash sum. The two are not the same number.

This number does not exist yet

Block length, B, the random seed, and the pass line: none of the four is locked yet
Until those four are written down, AV-DSR is empty
Empty is not written as 0, and no star is added

When the standard DSR is already empty (fewer than 30 fills, variance of 0, a price that is not finite, or a missing N), AV-DSR is empty too. After the calibration record writes those four items down and the regression test passes, this item replaces the DSR star. The replacement still reads only this user book.